Communication Networks

Transcription

1 Communication Networks Routing and Traffic Management Manuel P. Ricardo Faculdade de Engenharia da Universidade do Porto

2 Routing

3 Graph Directed and Undirected

4 Tree Trees T = (V,E)» graph with no cycles» number of edges E = V» any two vertices of the tree are connected by exactly one path A tree T is said to span a graph G = (V,E) (spanning tree) if» T = (V,E ) and E E

5 Shortest Path Trees Graphs and Trees can be weighted» G=(V,E,w)» T=(V, E,w)» w: ER Total cost of a tree T Minimum Spanning Tree T*» algorithms used to compute MST: Prism, Kruskal Shortest Path Tree (SPT) Rooted at Vertex s» tree composed by the» union of the shortest paths between s and each of other vertices of G» Algorithms used to compute SPT: Dijkstra, Bellman-Ford Computer networks use Shortest Path Trees

6 Routing in Datagram Networks

7 Forwarding, Routing Forwarding data plane» directing packet from input to output link» using a forwarding table Routing control plane» computing paths the packets will follow» routers exchanging messages» each router creating its forwarding table 0 destination address routing algorithm local forwarding table header value output link

8 Importance of Routing End-to-end performance» path affects quality of service» delay, throughput, packet loss Usage of network resources» balance traffic over routers and links» avoiding congestion by directing traffic to less-loaded links Transient disruptions» failures, maintenance» limiting packet loss and delay during changes

9 Shortest-Path Routing Path-selection model» Destination-based» Load-insensitive (e.g., static link weights)» Minimum hop count or minimum sum of link weights

10 Shortest-Path Problem Given a network topology with link costs» c(x,y) - link cost from node x to node y» Infinity - if x and y are not direct neighbors Compute the least-cost paths from source u to all nodes p(v) - node predecessor of node v in the path to u 2 u p(v) 4 3 v

11 Dijkstra s Shortest-Path Algorithm Iterative algorithm» After k iterations known least-cost paths to k nodes S set of nodes whose least-cost path is known» Initially, S={u}, where u is the source node» Add one node to S in each iteration D(v) current cost of path from source to node v» Initially D(v)=c(u,v) for all nodes v adjacent to u D(v)= for all other nodes v» Continually update D(v) when shorter paths are learned

12 Dijsktra s Algorithm Initialization: 2 S = {u} 3 for all nodes v 4 if v adjacent to u { 5 D(v) = c(u,v) } 6 else D(v) = 7 8 Loop 9 find node w not in S with the smallest D(w) 0 add w to S update D(v) for all v adjacent to w and not in S: 2 D(v) = min{d(v), D(w) + c(w,v)} 3 until all nodes in S

13 Dijkstra s Algorithm - Example

14 Dijkstra s Algorithm - Example

15 Shortest-Path Tree Shortest-path tree from u v 2 y 3 x 4 u 2 5 t w 4 3 s Forwarding table at u link z v (u,v) w (u,w) x (u,w) y (u,v) z (u,v) s (u,w) t (u,w)

16 Link-State Routing Each router keeps track of its incident links» link up, link down» cost on the link Each router broadcasts the link state every router gets a complete view of the graph Each router runs Dijkstra s algorithm, to» compute the shortest paths» construct the forwarding table Example protocols» Open Shortest Path First (OSPF)» Intermediate System Intermediate System (IS-IS)

17 Detection of Topology Changes Beacons generated by routers on links» Periodic hello messages in both directions» few missed hellos link failure hello 7

18 Broadcasting the Link State How to Flood the link state?» every node sends link-state information through adjacent links» next nodes forward that info to all links except the one where the information arrived X A C B D (a) X A C B D (b) When to initiate flooding?» Topology change link or node failure/recovery link cost change» Periodically refresh link-state information typically 30 minutes X A C B D (c) X A C B D (d)

19 Scaling Link-State Routing Overhead of link-state routing» flooding link-state packets throughout the network» running Dijkstra s shortest-path algorithm Introducing hierarchy through areas Area Area 2 area border router Area 0 Area 3 Area 4

20 Bellman-Ford Algorithm Define distances at each node x» d x (y) = cost of least-cost path from x to y Update distances based on neighbors» d x (y) = min {c(x,v) + d v (y)} over all neighbors v u 3 2 v w 4 2 x 5 s 3 4 y t z d u (z) = min{c(u,v) + d v (z), c(u,w) + d w (z)}

21 Distance Vector Algorithm c(x,v) = cost for direct link from x to v node x maintains costs of direct links c(x,v) D x (y) = estimate of least cost from x to y node x maintains distance vector D x = [D x (y): y є N ] Node x maintains also its neighbors distance vectors for each neighbor v, x maintains D v = [D v (y): y є N ] Each node v periodically sends D v to its neighbors» and neighbors update their own distance vectors» D x (y) min v {c(x,v) + D v (y)} for each node y N Over time, the distance vector D x converges

22 Distance Vector Algorithm Iterative, asynchronous each local iteration caused by: local link cost change distance vector update message from neighbor Distributed» node notifies neighbors only when its DV changes Each node: wait for (change in local link cost or message from neighbor) recompute estimates if DV to any destination has changed, notify neighbors Neighbors then notify their neighbors, if necessary

23 Distance Vector Example - Step 0 Optimum -hop paths Table for A Dst Cst Hop Table for B Dst Cst Hop E 3 C A 0 A B 4 B C D E 2 E F 6 F A 4 A B 0 B C D 3 D E F F 2 A 6 4 F B 3 D Table for C Table for D Table for E Table for F Dst Cst Hop Dst Cst Hop Dst Cst Hop Dst Cst Hop A A A 2 A A 6 A B B 3 B B B B C 0 C C C C C C D D D 0 D D D E E E 0 E E 3 E F F F F 3 F F 0 F

24 Distance Vector Example - Step Optimum 2-hop paths Table for A Dst Cst Hop Table for B Dst Cst Hop E 3 C A 0 A B 4 B C 7 F D 7 B E 2 E F 5 E A 4 A B 0 B C 2 F D 3 D E 4 F F F 2 A 6 4 F B 3 D Table for C Table for D Table for E Table for F Dst Cst Hop Dst Cst Hop Dst Cst Hop Dst Cst Hop A 7 F A 7 B A 2 A A 5 B B 2 F B 3 B B 4 F B B C 0 C C C C 4 F C C D D D 0 D D D 2 C E 4 F E E 0 E E 3 E F F F 2 C F 3 F F 0 F

25 Distance Vector Example - Step 2 Optimum 3-hop paths Table for A Dst Cst Hop Table for B Dst Cst Hop E 3 C A 0 A B 4 B C 6 E D 7 B E 2 E F 5 E A 4 A B 0 B C 2 F D 3 D E 4 F F F 2 A 6 4 F B 3 D Table for C Table for D Table for E Table for F Dst Cst Hop Dst Cst Hop Dst Cst Hop Dst Cst Hop A 6 F A 7 B A 2 A A 5 B B 2 F B 3 B B 4 F B B C 0 C C C C 4 F C C D D D 0 D D 5 F D 2 C E 4 F E 5 C E 0 E E 3 E F F F 2 C F 3 F F 0 F

26 Maximum Flow of a Network

27 Flow Network Model Flow network» source s» sink t» nodes a, b and c Edges are labeled with capacities» (e.g. bit/s) Communication networks are not flow networks» they are queue networks» flow networks enable to determine limit values

28 Maximum Capacity of a Flow Network Max-flow min-cut theorem» maximum amount of flow transferable through a network» equals minimum value among all simple cuts of the network Cut split of the nodes V into two disjoint sets S and T» S T = V» there are 2 V 2 possible cuts Capacity of cut (S, T):

29 Max-flow Min-cut - Example possible cuts Maximum flow = 0

30 Traffic Management Packet Level, Flow Level

31 Time Scales Packet Level» Queueing & scheduling at multiplexing points» Determines performance offered to packets over a short time scale (ms) Flow Level» Management of traffic flows & resource allocation ( ms to s)» Matching traffic flows to available resources Flow-Aggregate Level» Routing of aggregate traffic flows across the network (min to days)» Efficient utilization of resources; meeting of service levels» Traffic Engineering

32 End-to-End QoS Packet traversing network» encounters delay and possible loss» at various multiplexing points Packet buffer 2 N N End-to-end performance is the accumulation of per-hop performances

33 Quality of Service and Scheduling End-to-End QoS» Packet delay, packet loss» Performance level can be engineered by Buffer control, bandwidth control Admission control of traffic entering network Scheduling Concepts: fairness, priority Fair Queueing» Weighted Fair Queuing (WFQ)» Packetized Generalized Processor Sharing (PGPS) Guaranteed Service: WFQ, Rate-control

34 FIFO Queueing Arriving packets Packet discard when full Packet buffer Transmission link All packet flows share the same buffer Transmission Discipline: First-In, First-Out Buffering Discipline» Discard arriving packets if buffer is full» Random discard» Discard head-of-line (i.e. oldest, packet)

35 FIFO Queueing Cannot provide differential QoS to different packet flows Statistical delay guarantees via load control» Restrict number of flows allowed (connection admission control)» Difficult to determine performance delivered Finite buffer determines a maximum possible delay Buffer size determines loss probability» Depends on arrival & packet length statistics

36 FIFO Queueing with Discard Priority (a) Arriving packets Packet discard when full Packet buffer Transmission link (b) Arriving packets Class discard when full Packet buffer Class 2 discard when threshold exceeded Transmission link

37 Head Of Line (HOL) Priority Queueing Packet discard when full High-priority packets Transmission link Low-priority packets Packet discard when full When high-priority queue empty High priority queue serviced until empty High priority queue has lower waiting time Buffers can be dimensioned for different loss probabilities Surge in high priority queue can cause low priority queue to saturate

38 Delay HOL Priority Features Provides differential QoS High-priority classes can» get all the bandwidth» starve lower priority classes Per-class loads

39 Earliest Due Date Scheduling Arriving packets Tagging unit Sorted packet buffer Packet discard when full Transmission link Queue in order of due date» packets requiring low delay get earlier due date» packets without delay get indefinite or very long due dates

40 Fair Queueing / Generalized Processor Sharing Packet flow Packet flow 2 Approximated bit-level round robin service C bit/s Packet flow n Transmission link Each flow has its own logical queue» prevents hogging; allows differential loss probabilities C bit/s allocated equally among non-empty queues» transmission rate = C / n(t), where n(t)= # non-empty queues Idealized system assumes fluid flow from queues Implementation requires approximation» simulate fluid system; sort packets according to completion time in ideal system

41 Fluid Flow System Buffer at t=0 Buffer 2 at t=0 0 2 t Fluid-flow system: both packets served at rate /2 Both packets complete service at t = 2 Packet from buffer 2 waiting Packet-by-packet system: buffer served first at rate ; then buffer 2 served at rate. Packet from buffer 2 being served Packet from buffer being served 0 2 t

42 Fluid Flow Buffer at t=0 Buffer 2 at t=0 2 Fluid-flow system: both packets served at rate /2 Packet from buffer 2 served at rate t Packet from buffer 2 waiting Packet-by-packet fair queueing: buffer 2 served at rate Packet from buffer served at rate t

43 Fluid Flow with Weights Buffer at t=0 /4 Fluid-flow system: packet from buffer served at rate /4; Buffer 2 3/4 at t=0 Packet from buffer served at rate Packet from buffer 2 served at rate 3/4 0 2 t Packet from buffer waiting Packet-by-packet weighted fair queueing: buffer 2 served first at rate ; then buffer served at rate Packet from buffer served at rate Packet from buffer 2 served at rate 0 2 t

44 Packetized GPS/WFQ Arriving packets Tagging unit Sorted packet buffer Packet discard when full Transmission link Compute packet completion time in ideal system» add tag to packet» sort packet in queue according to tag» serve according to HOL

45 Bit-by-Bit Fair Queueing Assume n flows, n queues round = cycle serving all n queues If each queue gets bit per cycle Round number = number of cycles of service that have been completed rounds Current Round # If packet arrives to idle queue: Finishing time = round number + packet size in bits If packet arrives to active queue: Finishing time = finishing time of last packet in queue + packet size

46 Number of rounds = Number of bit transmission opportunities Rounds Buffer Buffer 2 Buffer n Packet completes transmission k rounds later Packet of length k bits begins transmission in this time Differential Service» If a traffic flow is to receive twice as much bandwidth as a regular flow, then its packet completion time would be half

47 Computing the Finishing Time F(i,k,t) = finish time of kth packet that arrives at time t to flow i P(i,k,t) = size of kth packet that arrives at time t to flow i R(t) = round number at time t rounds Generalize so R(t) continuous, not discrete R(t) grows at rate inversely proportional to n(t) Fair Queueing:(take care of both idle and active cases) F(i,k,t) = max{f(i,k-,t), R(t)} + P(i,k,t) Weighted Fair Queueing: F(i,k,t) = max{f(i,k-,t), R(t)} + P(i,k,t)/w i

48 Traffic Management at the Flow Level

49 Network Congestion Congestion 3 6 Congestion occurs when a surge of traffic overloads network resources Congestion Control» Preventive: Scheduling & Reservations» Reactive: Detect & Discard

50 Throughput Ideal Congestion Control Ideal effect of congestion control: resources used efficiently up to capacity available Controlled Uncontrolled Offered load

51 Congestion Control Open loop Closed loop

52 Open-Loop Control Network performance is guaranteed to all traffic flows that have been admitted into the network Initially for connection-oriented networks Key Mechanisms» Admission Control» Policing» Traffic Shaping» Traffic Scheduling

53 Bit/s Admission Control Flow negotiates contract with network Peak rate Average rate Specify requirements» Offered traffic (traffic contract) Peak, Average, Min Bit rate Maximum burst size» QoS required Delay, Loss Typical bit rate demanded by a variable bit rate information source Time Network computes resources needed» Effective bandwidth If flow is accepted: network allocates resources to ensure QoS as long as source conforms to contract

54 Policing Network monitors traffic flow continuously to ensure it meets traffic contract When a packet violates the contract network can discard or tag the packet giving it lower priority If congestion occurs tagged packets are discarded first Leaky Bucket Algorithm» commonly used policing mechanism» Bucket has specified leak rate for average contracted rate: r depth to accommodate variations in arrival rate: B» Arriving packet is conforming if it does not result in bucket overflow

55 Leaky Bucket Algorithm Leaky Bucket algorithm can be used to police arrival rate of a packet stream water poured irregularly leaky bucket water drains at a constant rate Leak rate corresponds to longterm rate Bucket depth corresponds to maximum allowable burst arrival packet per unit time Assume constant-length packet Let X = bucket content at last conforming packet arrival Let t a last conforming packet arrival time = depletion in bucket

56 Leaky Bucket Example I = 4, B=0, L=B-I=6 Packet arrival Nonconforming Time B Bucket content I * * * * * * * * * Time Non-conforming packets not allowed into bucket, hence not included in calculations

57 Leaky Bucket Algorithm Arrival of a packet at time t a X = X - (t a - LCT) Depletion rate: packet per unit time Current bucket content X < 0? Interarrival time Yes L=B-I I = increment per arrival, nominal interarrival time Non-empty No X = 0 Nonconforming packet arriving packet would cause overflow Yes X > L? No X = X + I LCT = t a conforming packet empty X = value of the leaky bucket counter X = auxiliary variable LCT = last conformance time conforming packet

58 Dual Leaky Bucket Dual leaky bucket to police PCR, SCR, and MBS: Incoming traffic Leaky bucket SCR and MBS Tagged or dropped Untagged traffic Leaky bucket 2 PCR and CDVT Tagged or dropped Untagged traffic PCR = peak cell rate CDVT = cell delay variation tolerance SCR = sustainable cell rate MBS = maximum burst size

59 Traffic Shaping Traffic shaping Policing Traffic shaping Policing Network A Network B Network C Networks police the incoming traffic flow Traffic shaping is used to ensure that a packet stream conforms to specific parameters Networks can shape their traffic prior to passing it to another network

60 Leaky Bucket Traffic Shaper Incoming traffic Size N Server Shaped traffic Packet Buffer incoming packets Play out periodically to conform to parameters Surges in arrivals are buffered & smoothed out Possible packet loss due to buffer overflow Too restrictive, since» conforming traffic does not need to be completely smooth

61 Leaky Bucket Traffic Shaper Controls the maximum instantaneous rate (leak rate) r of a flow» accepts bursty traffic on its input» eliminates bursts on the output While the buffer is not empty» the output flow is periodic (T = / r),» with rate r Maximum accepted Burst Size» MBS = B * R / (R r) for R > r

62 Token Bucket Traffic Shaper An incoming packet must have sufficient tokens before admission into the network Token Tokens arrive periodically ( r ) Size b Incoming traffic Size N Server Shaped traffic Packet Token rate regulates transfer of packets If sufficient tokens available, packets enter network without delay b determines how much burstiness allowed into the network

63 Token Bucket Shaping Effect The token bucket constrains the traffic from a source to be limited to b + r t bits in an interval of length t b bytes instantly b + r t r bytes/second t

64 Packet transfer with Delay Guarantees (a) A(t) = b+rt Token Shaper Bit rate > R > r e.g., using WFQ R(t) No backlog of packets R(t) (b) Buffer occupancy at 0 b Assume fluid flow for information Token bucket shaper allows burst of b then r byte/s» Since R>r, buffer never greater than b byte» Thus, mux < b/r Rate into second mux is r<r, so bytes are never delayed b R 2 b R - r Buffer occupancy at 2 t Empty t

65 Delay Bounds with WFQ / PGPS Assume» traffic shaped (token bucket) to parameters b & r» schedulers give flow at least rate R>r» H hop path» m is maximum packet size for the given flow» M maximum packet size in the network» R j transmission rate in jth hop Maximum end-to-end delay experienced by a packet from flow i is D b R ( H ) m R H j M R j

66 Scheduling for Guaranteed Service Suppose guaranteed bounds on end-to-end delay across the network are to be provided. Then» Call admission control procedure is required to allocate resources & set schedulers» Traffic flows from sources must be shaped/regulated so that they do not exceed their allocated resources Strict delay bounds can be met

67 Current View of Router Functions Routing Agent Reservation Agent Mgmt. Agent Admission Control [Routing database] [Traffic control database] Input driver Classifier Internet forwarder Pkt. scheduler Output driver

68 Closed-Loop Control Congestion control» Feedback information to regulate flow from sources into network» Based on buffer content, link utilization,» Examples: TCP at transport layer; congestion control at ATM level End-to-end vs. Hop-by-hop» Delay in effecting control Implicit vs. Explicit Feedback» Source deduces congestion from observed behavior» Routers/switches generate messages alerting to congestion

69 End-to-End vs. Hop-by-Hop Congestion Control Source Packet flow Destination (a) Source Destination (b) Feedback information

70 TCP Congestion Control End-to-end, Implicit Main idea» each source determines its capacity» based on criteria enabling flow fairness efficiency Received ACKs regulate packet transmission they are the source clock

71 Additive Increase/Multiplicative Decrease Changes in channel capacity adjustment of transmission rate New variable per connection CongestionWindow Bitrate (byte/s) CongestionWindow/RTT Objective» If network congestion decreases CongestionWindow increases» If network congestion increases CongestionWindow decreases

72 Additive Increase/Multiplicative Decrease How does the source know if/when the network is in congestion? By timeout!» wired link low BER low FER» timeout occurrence loss of packet» packet loss buffer in router is full congestion

73 Additive Increase/Multiplicative Decrease Algorithm» increases CongestionWindow by segment for each RTT (Round Trip Time) additive increase Source Destination» divide CongestionWindow by 2 when there is a packet loss multiplicative decrease

74 KB Additive Increase/Multiplicative Decrease Saw-tooth behavior T ime (seconds)

75 Slow Start Objective» determine the available capacity Source Destination Behaviour» start by CongestionWindow = segment» double CongestionWindow by each RTT

76 Fast Retransmission, Fast Recovery Problem» if TCP timeout is large long inactivity period Sender Packet Packet 2 Packet 3 Packet 4 Receiver ACK ACK 2 Packet 5 ACK 2 Solution Packet 6 ACK 2» fast retransmission ACK 2 after 3 repeated ACKs Retransmit packet 3 ACK 6

77 Source Destination TCP Slow Start Slow Start» Sender starts with CongestionWindow=sgm» Doubles CongestionWindow by RTT When a segment loss is detected, by timeout» threshold = ½ congestionwindow(*)» CongestionWindow= sgm (router gets time to empty queues)» Lost packet is retransmitted» Slow start while congwindow<threshold» Then Congestion Avoidance phase (*) - in fact FlightSize, the amount of outstanding data

78 KB Congestion Avoidance Congestion Avoidance (additive increase)» increments congestionwindow by sgm, per RTT Detection of segment loss, by reception of 3 duplicated ACKs» Assumes packet is lost, Not by severe congestion, because following segments have arrived» Retransmits lost packet» CongestionWindow=CongestionWindow/2» Congestion Avoidance phase Source Destination T ime (seconds) 0.0

79 Desirable Bandwidth Allocation Max-min fairness Fair use gives bandwidth to all flows (no starvation)» Max-min fairness gives equal shares of bottleneck Bottleneck link max-min fair: maximizes the minimum bandwidth flows get

80 Desirable Bandwidth Allocation Bitrates along the time Bitrates must converge quickly when traffic patterns change Flow slows quickly when Flow 2 starts Flow speeds up quickly when Flow 2 stops CN5E by Tanenbaum & Wetherall, Pearson Education-Prentice Hall and D. Wetherall, 20

81 Homework Review slides Read from Leon-Garcia» Chapter 7 Packet Switching Networks Read from Tanenbaum» TCP Congestion control For formal analysis of these topics» Routing: Bertsekas & Gallager, Data Networks» Traffic and resource management: Anurag Kumar, D. Manjunath, Joy Kuri, Communication Networking - An Analytical Approach, Elsevier, 2004